Erdos-Renyi generator had bad logic in thrust calls (#4362)

The `random_iterator`, a thrust transform iterator that was used for counting and filtering returns the probability, but they `copy_if` call was assuming it was returning the index.

Modified the logic and then consolidated the `copy_if` and `transform` into a single call with an output iterator.

Closes #4359

Authors:
  - Chuck Hastings (https://github.com/ChuckHastings)

Approvers:
  - Seunghwa Kang (https://github.com/seunghwak)
  - Naim (https://github.com/naimnv)
  - Joseph Nke (https://github.com/jnke2016)

URL: #4362

Author: ChuckHastings

cpp/src/generators/erdos_renyi_generator.cu

@@ -23,10 +23,9 @@
 #include <thrust/copy.h>
 #include <thrust/count.h>
 #include <thrust/iterator/counting_iterator.h>
-#include <thrust/iterator/transform_iterator.h>
+#include <thrust/iterator/transform_output_iterator.h>
 #include <thrust/iterator/zip_iterator.h>
 #include <thrust/random.h>
-#include <thrust/transform.h>
 #include <thrust/tuple.h>
 
 namespace cugraph {
@@ -42,45 +41,38 @@ generate_erdos_renyi_graph_edgelist_gnp(raft::handle_t const& handle,
   CUGRAPH_EXPECTS(num_vertices < std::numeric_limits<int32_t>::max(),
                   "Implementation cannot support specified value");
 
-  auto random_iterator = thrust::make_transform_iterator(
-    thrust::make_counting_iterator<size_t>(0),
-    cuda::proclaim_return_type<float>([seed] __device__(size_t index) {
-      thrust::default_random_engine rng(seed);
-      thrust::uniform_real_distribution<float> dist(0.0, 1.0);
-      rng.discard(index);
-      return dist(rng);
-    }));
+  size_t max_num_edges = static_cast<size_t>(num_vertices) * num_vertices;
 
-  size_t count = thrust::count_if(handle.get_thrust_policy(),
-                                  random_iterator,
-                                  random_iterator + num_vertices * num_vertices,
-                                  [p] __device__(float prob) { return prob < p; });
-
-  rmm::device_uvector<size_t> indices_v(count, handle.get_stream());
+  auto generate_random_value = cuda::proclaim_return_type<float>([seed] __device__(size_t index) {
+    thrust::default_random_engine rng(seed);
+    thrust::uniform_real_distribution<float> dist(0.0, 1.0);
+    rng.discard(index);
+    return dist(rng);
+  });
 
-  thrust::copy_if(handle.get_thrust_policy(),
-                  random_iterator,
-                  random_iterator + num_vertices * num_vertices,
-                  indices_v.begin(),
-                  [p] __device__(float prob) { return prob < p; });
+  size_t count = thrust::count_if(handle.get_thrust_policy(),
+                                  thrust::make_counting_iterator<size_t>(0),
+                                  thrust::make_counting_iterator<size_t>(max_num_edges),
+                                  [generate_random_value, p] __device__(size_t index) {
+                                    return generate_random_value(index) < p;
+                                  });
 
   rmm::device_uvector<vertex_t> src_v(count, handle.get_stream());
   rmm::device_uvector<vertex_t> dst_v(count, handle.get_stream());
 
-  thrust::transform(handle.get_thrust_policy(),
-                    indices_v.begin(),
-                    indices_v.end(),
-                    thrust::make_zip_iterator(thrust::make_tuple(src_v.begin(), src_v.end())),
+  thrust::copy_if(handle.get_thrust_policy(),
+                  thrust::make_counting_iterator<size_t>(0),
+                  thrust::make_counting_iterator<size_t>(max_num_edges),
+                  thrust::make_transform_output_iterator(
+                    thrust::make_zip_iterator(src_v.begin(), dst_v.begin()),
                     cuda::proclaim_return_type<thrust::tuple<vertex_t, vertex_t>>(
                       [num_vertices] __device__(size_t index) {
-                        size_t src = index / num_vertices;
-                        size_t dst = index % num_vertices;
-
-                        return thrust::make_tuple(static_cast<vertex_t>(src),
-                                                  static_cast<vertex_t>(dst));
-                      }));
-
-  handle.sync_stream();
+                        return thrust::make_tuple(static_cast<vertex_t>(index / num_vertices),
+                                                  static_cast<vertex_t>(index % num_vertices));
+                      })),
+                  [generate_random_value, p] __device__(size_t index) {
+                    return generate_random_value(index) < p;
+                  });
 
   return std::make_tuple(std::move(src_v), std::move(dst_v));
 }

cpp/tests/generators/erdos_renyi_test.cpp

@@ -87,6 +87,7 @@ void er_test(size_t num_vertices, float p)
 TEST_F(GenerateErdosRenyiTest, ERTest)
 {
   er_test<int32_t>(size_t{10}, float{0.1});
+  er_test<int32_t>(size_t{10}, float{0.5});
   er_test<int32_t>(size_t{20}, float{0.1});
   er_test<int32_t>(size_t{50}, float{0.1});
   er_test<int32_t>(size_t{10000}, float{0.1});